As the second decade of twenty first century comes to and end, it is not far from truth to name it the “Decade of AI Maturity”. The emergence of deep neural networks, development of innovative and powerful network architectures such as Generative Adversarial Networks (GANs), the extensive use of parallel computing through GPUs and a the giant leap forward in Reinforcement Learning toward winning the game of Go, Warcraft and self driving cars all have happened in this decade. Now with this promising upward trend of improvements, are we set to reach a level of AI developments that would result into building SkyNet and Terminators?

Open AI recently published a blog post on how their agents learned to play “Hide-and-Seek” through Reinforcement Learning. Having autonomous agents playing hide and seek seems to be such a fun and innocent outcome but the details of what these agents learned throughout the training sheds light on the power of AI. Based on Open AI article, the agents were trained by playing the game for over 470 million iterations.

Reinforcement Learning is a branch of machine learning in which the model learns from the feedbacks, in form of rewards, it receives for its actions. The ultimate goal of the model is to maximize the amount of reward it receives and throughout trying and error, it will finally learn how to perform the action at hand. For example in playing chess, the reward would come at the end of the game if the player wins and it would be propagated back to all the actions taken throughout the game and with repeating the game so many times the model will learn how to play the game.

Back to OpenAI case, the concept is the same for hide-and-seek agents. The goal for them was to try to win the game by any means. This had led them to develop unexpected skillsets throughout the training. The agents gradually learned to tweak the environment around them toward their leverage. From building shelter using the objects, to use ramps to jump over the walls, they managed to get smarter and smarter.

Now the question is, what type of boundaries should we define for these type of agents? One might challenge that having these “virtual” agents playing a “specific task”, in this case hide-and-seek, in a confined “ virtual environment” is not concerning as it is in virtual world and it is also an innocent game.

But in my point of view, it needs our attention. First of all we are not very close to implement such intelligent agents in physical assets. Just couple weeks ago Boston Dynamics published yet another jaw dropping video of its humanoid robot performing Parkour. Proving that not only the mechanical design of the robot endures various types of shear, impact and tension associated with performing these actions but also their control software is perfectly capable of analyzing the environment, identifying obstacles and control the actuators to perform the movements. This or any similar technology provides suitable physical capabilities on perform actions in real world. Soon we can see Boston Dynamics Atlas robots moving boxes around to make shelter or climb ladders to jump into the other team’s shelter while playing hide and seek.

Secondly, in OpenAI research we can observe how these agents came of with innovative solutions given very few available options in their environment. If we extrapolate those innovations to a much more complex environment and a different task, we then it makes think if we need to have a strict review and screening process for them to make sure in the physical world, they wouldn’t perform any harmful action? I think we must. Actually while I was drafting this article, I read through the most recent issue of deeplearning.ai’s “The Batch” in which Andrew Ng discusses similar concerns. Here is short quote from it:

I’d highly suggest you’d subscribe to The Batch to get the weekly news on AI.

Back to our discussion about the threat of unpredicted smart actions by robots, fortunately once the virtual agents are moved to their physical hosts, they will be constrained by time just like us, humans, and they won’t be able to improve as fast as it was possible in the virtual realm (we should remember they were able to achieve such skills in 470 million iterations in virtual world which is not quite practical in physical world) but then there is alway potential for them to scan the environment, pass the information to virtual world, perform training and upgrade their intelligence.

So should we start thinking about mandating certain ethics for AI, I believe Yes. Although we are still far from reaching Artificial General Intelligence (AGI) but it is time to start thinking about an ethical foundation for AI. how do we do it and what measures we should consider is open for discussion.

Leave your thoughts in the comments section below so we can discuss.